How To Make AI Work For You (Not The Other Way Around)

Like any CX technology, AI should work for you, not the other way around. On the one hand, it shouldn’t be left unchecked. You need to train it and watch as it works. But you don’t want to be constantly programming it either (you might as well adopt a decision-tree-based chatbot if you want to do that). That’s not saving time — it’s adding to your team’s workload.

Managing AI is about finding the balance. With the proper training and oversight, you want it to handle repetitive tasks and improve on its own. These five tips for working with AI will ensure it becomes your partner, not your problem.

1. Integrate AI with the right tools for context

AI is only as good as the information it’s given. This is especially true when interacting with customers who have a much broader experience with your company than the message they’ve most recently sent.

Say a SaaS company gets a chat message from a customer having a technical issue with their tool. The AI powering that chatbot is connected to their knowledge base and responds with accurate instructions on correcting the issue. But it’s not connected to the help desk where your team handles issues over email, so it misses that this customer has previously reported that those instructions don’t work.

The customer feels frustrated and demands to speak with a person, having lost confidence that AI can help them. The more context you provide AI, the better. If you’re using automated agents, consider integrating AI technologies with customer relationship management (CRM), enterprise resource planning (ERP), shipping, marketing automation, and customer service tools.

Say you use Zendesk to host your help center and want your AI-powered automated agent to be able to send customers articles directly. You can integrate Zendesk with Forethought so AI can provide direct links and summarize them accurately for customers.

These integrations will help you spend less time responding to tickets AI couldn’t close and more time managing integrations that keep it running like clockwork. Forethought integrates with over 40 of the most common tools companies use worldwide, including Zendesk, HubSpot, Salesforce, Amazon Web Services (AWS), and Dialpad.

2. Help AI understand your knowledge base

Both customers and AI appreciate when a knowledge base is easy to navigate. Whether you’re using it as a source of information for an automated agent or expecting AI to send articles directly to customers, it’s got to be up-to-date and easy to understand.

Without a good source of information, AI might give customers confusing responses or, worse, send them to articles that don’t help. Your team will end up responding to those tickets instead of managing the source of information that helps AI run smoothly. Some of the most important best practices to consider when creating and organizing a knowledge base focus on clear language and concise explanations:

• Use the words customers use. If your customers say “reset password” and you say “account retrieval,” use “reset password.”
• Create one article per specific use case. Maybe “account retrieval” covers both “reset password” and “change password.” Create separate articles for each.
• Use clear and concise titles. A title like “Resolving issues related to entering your account” is terrible. “Login Issues” would make more sense. 

Etekcity’s help center follows these examples with clear titles, dedicated articles for each specific issue, and very concise instructions for their customers.

Etekcity’s knowledge base is one information source for Solve, their AI-powered chatbot. Along with previously successful tickets and other data sources, their knowledge base contributed to a 48% self-service rate and a 3x increase in ticket resolution when they implemented AI.

3. Train your AI’s intents

Intents are important to AI and natural language understanding (NLU). They are the backbone of the machine learning models that allow AI to understand and interpret human language. They explain the purpose or point of a customer’s question or request.

An automated AI agent designed for customer service might have many intents, which developers can use to train the system to recognize and provide the right response. Here are some example intents and the training phrases developers might use to help recognize them:

• Change a gym membership
  • I want to switch to the premium plan for access to all locations.
  • How can I pause my membership for two months while I recover from surgery?
  • Can I downgrade to the basic membership? I don’t use the group classes.
  • I want to add a personal trainer to my membership plan. How do I do that?
  • Can I extend my membership renewal by a month due to travel?
• Replace a damaged furniture order
  • My table arrived with a scratch. Can you replace it?
  • The couch I ordered has a broken leg. How do I report this?
  • Can you send me a replacement for the damaged bookshelf I just received?
  • What’s the process for exchanging the dresser that arrived with a cracked mirror?
  • The dining set I bought was missing a chair. How can I get one?
• Reschedule dog grooming appointment
  • Can I move my grooming appointment to next Thursday afternoon?
  • I need to cancel today’s appointment. Can I book a new slot for next week?
  • I’m running late. Can the groomer wait 15 minutes, or do I need to reschedule?
  • How do I reschedule my dog’s appointment for a time when I’m off work?
  • Can I change the appointment to a full grooming session instead of just a bath?

If you’re using Forethought Solve, you can train intents for chat and email directly in the Workflow builder.

Each intent you create should have a particular purpose and be grounded in common customer language your support team has collected. Make sure to provide a diverse set of training phrases so that your AI can pick up on changes in language and context.

If your intents aren’t diverse, you might be unable to deflect as many issues as you’d hoped. AI will only pick up how to respond in a few specific situations, and your team gets stuck fielding tickets it couldn’t solve instead of proactively training its model.

4. Refine agentic AI policies

If intents are the backbone of how AI understands human language, policies are how agentic AI understands how customers want them to act.

Autoflows is an agentic AI solution for teams using Solve. It allows automated agents to not only respond to customers but take action to resolve their issues. It’s powered by policies that are automatically built using AI but can be tweaked when needed. If you focus on getting your policies right, you’ll spend less time reacting to issues it couldn’t deflect.

Your policies will guide AI on the backend of your tool, where you can review AI’s actions and optimize them if you spot room for improvement.

Here’s what a policy might look like for a customer who wants to pause their subscription to a SaaS tool:

If a customer mentions pausing a subscription:

1. Reassure them that pausing their account is possible and won’t result in data loss
2. Ask them how long they would like to pause the subscription (month, year, custom period)
3. If monthly, inform them the pause can last for up to three months
4. If annual, inform them the pause will extend the subscription period at the end of the year
5. Provide a link to their account management page
6. Ask them to log in and select the “Pause Subscription” option
7. Once they’ve selected the pause period, confirm with the customer:
   “Your subscription will be paused from [start date] to [end date]. Does this work for you?”
8. If they say yes, confirm the action in the system and send a confirmation email
9. If they say no, adjust the pause period as needed or offer them the option to contact support
10. If they can’t find the pause option, guide them through the process or offer a handoff
11. If the pause is not supported, apologize and suggest a cancellation

Here’s what a policy might look like for a customer who wants to file a warranty claim for a broken laptop: 

If a customer mentions a broken device and a warranty claim

1. Start by empathizing:
   “I’m sorry to hear about the issue with your laptop. Let me help you with the warranty.”
2. Request the model number, serial number, and proof of purchase (e.g., invoice or receipt)
3. Provide a link for uploading these documents
4. Check if the laptop is within the warranty period using the purchase date and model number
5. If eligible, inform the customer:
   “Your laptop qualifies for warranty repair or replacement. Let’s proceed.”
6. If not eligible, explain why and provide alternative options like paid repair services
7. If repair is needed, provide the shipping address for sending the laptop
8. Generate a prepaid shipping label
9. If replacement is required, confirm the shipping address and expected delivery date
10. If they can’t find the model number or receipt, offer a step-by-step guide for finding them
11. If the laptop damage does not qualify, explain the terms and conditions of the warranty

Here’s what a policy might look like for a customer who is missing items from their grocery order:

If the customer reports missing items

1. Apologize for the inconvenience and reassure them
   “I’m sorry about the missing items in your order. Let me sort this out for you right away.”
2. Ask for the order number and delivery date
3. Confirm the list of items in the order and the items the customer reports missing
4. If the missing items are verified as part of the order, offer a refund or a redelivery option
5. If offering a refund, inform the customer
   “I’ll refund the missing items. You’ll see the refund within 3-5 business days.”
6. If offering redelivery, confirm their delivery address and schedule a redelivery time
7. If the customer reports perishables as missing, prioritize redelivery within the same day
8. If the customer received the wrong items, arrange for the correct items to be delivered and provide instructions for disposing of or donating the incorrect items
9. After resolving the issue, ask: “Is there anything else I can assist you with?”
10. Ensure a confirmation email or notification is sent detailing the resolution
11. If the customer can’t provide an order ID or the issue involves recurring errors, escalate to the “Customer Service Support” intent for further assistance.

You shouldn’t write these policies from scratch—let AI tools, like Forethought Discover, analyze your data and customer service interactions to make suggestions first. Regularly review them for clarity and complexity—the longer and more complex the policy, the greater the chance of AI getting confused.

Much like your intents, your policies should use conversational language. But there’s no need for excessive direction in tone. Instead, use condition if/then statements to guide the conversation generally.

3. Train AI for precise routing

NLU also allows you to route tickets quickly and accurately. Your goal is to ensure that it understands every issue well so that tickets get to the right place.

You’ll use a combination of ticket labels and descriptions to do so, but just like policies, you won’t have to start from Scratch. Triage, Forethought’s AI-powered routing technology, offers Triage Quickstart. It uses a combination of pre-built and custom models to automatically suggest fields based on your data.

Those pre-built models suggest fields based on sentiment, sentiment intensity, spam, and language, and then Triage’s language learning models (LLMs) refine them based on your historical interactions. Your goal is to focus on refining those labels and descriptions instead of reactively re-routing tickets that might not have made it to the right destination.

Here’s an example of good and bad labels for routing tickets after a damaged order.

The good examples link the issue to delivery, offer specific phrases for AI to recognize, and clearly outline exclusions. The bad examples are too generic, and the particular list of terms might lead to over-predictions.

Here’s an example for a customer who wants to cancel an online subscription:

The good example differentiates this label from other cancellation problems and offers examples to guide AI. The bad example is too broad, uses vague language, and attempts to account for too many edge cases, which could be confusing.

You ultimately want to train AI with focused, clear examples and exclusions. But if you’re too broad or too complex, you might find yourself with tickets routed to the wrong place. Some best practices include:

• A human-readable name or title.
• Clear description that accurately describes its purpose
• Use examples or notes when it is appropriate
• Refine the LLM definition in the description
• Explain why the label is important.
• Use clear formatting
• Focus on positive Instructions exclusions

It’s important to test your fields on a diverse set of cases. If you focus on just 10, for example, you might end up with 100% accuracy for those 10 but overfit for other scenarios.

5. Regularly review AI’s performance

The bulk of your energy when working with AI should go towards reviewing how it’s working and making improvements. A little time spent will mean AI stays helpful and evolves as your needs change. These are some of the most impactful ways to spend your time:

• Fill in AI’s knowledge gaps. AI depends on good information to do its job. Tools like Discover can find where your knowledge base is missing content and even draft articles to fill those gaps.
• Create smart intents. Using Discover’s two-level taxonomy, you can break down big topics like “Account Management” into detailed subtopics like “Password Reset.”
• Pay attention to customer sentiment. A sentiment analysis can show you why customers are frustrated. Those intents need attention if you notice repeat issues or dropping sentiment scores.
• Keep training data fresh. Customer language changes over time. As customer needs shift, regularly add new phrases to your intents, policies, and field descriptions.
• Double-check AI-generated content. Review AI-generated intents, policies, workflows, and fields before they go live. It’s better to catch errors upfront than clean up after them later.

If you spend time refining AI instead of reacting to it, you’ll keep it working for you—not the other way around. Make small adjustments regularly to keep everything running smoothly.

AI and humans should work together

A poorly managed AI system can create confusion and extra work. A well-managed one reduces repetitive tasks, supports your team, and evolves based on your guidance. Your role isn’t to fix its errors or handle what it couldn’t. It’s to guide, refine, and ensure the system aligns with your goals.

AI works best when humans stay in control. Your leadership and oversight make the difference between technology that just operated and one that truly works. AI shouldn’t replace your support team — there should be a partnership where AI and your team complement each other’s strengths.